Agent Evidence Recorder : Proof of what your AI agent did, redacted by default

Hi Product Hunt 👋 I built AER after a simple question I couldn't answer: an AI agent deleted a customer's data on request — can you prove what it actually did? Logging tools answer "was it fast, did it error." They don't answer "who asked for this, which tools ran, did a human approve the irreversible step, and was the PII handled safely." That's the gap. Agent Evidence Recorder records every critical step of an agent run as evidence: • A chronological timeline — user input, model calls, tool calls, human approvals, errors, final output. • Redacted by default — detected PII is never stored in the clear, only a hash and a redaction finding. Raw view exists, but it's an explicit, deliberate action. • Tamper-evident — events form an append-only hash chain, and every export carries a sha256 content hash. You can verify a run outside the system. • One portable JSON audit packet you can hand to a compliance or legal reviewer. Integration is a small SDK. Roughly: const run = await startRun({ apiKey, baseUrl }); await run.event({ type: "tool_call", name: "delete_customer" }); await run.complete({ status: "success" }); The repo ships a working demo: a Customer Data Deletion Request agent that hits every case — PII redaction, human approval before the irreversible step, and a completed evidence packet at the end. What it is not: an observability dashboard. No token-cost charts, no latency gauges. It answers one question — can we prove what the agent did, safely? It's early and I'd genuinely like to hear where this breaks for you: – Would your auditor accept a JSON packet, or do they need PDF? – What would you need before you'd trust this in a regulated workflow? Happy to answer anything today.